Trustworthy Misinformation Mitigation with Soft Information Nudging

Trustworthy misinformation mitigation with soft information nudging

Benjamin D. Horne Maur´ıcio Gruppi Sibel Adalı Department of Computer Science Department of Computer Science Department of Computer Science Rensselaer Polytechnic Institute Rensselaer Polytechnic Institute Rensselaer Polytechnic Institute Troy, NY, USA Troy, NY, USA Troy, NY, USA [email protected] [email protected] [email protected]

Abstract—Research in combating misinformation reports up the information’s correctness to debate which can lead many negative results: facts may not change minds, especially to suppression of minority voices and opinions. Furthermore, if they come from sources that are not trusted. Individuals can sophisticated misinformation campaigns mix correct and incor- disregard and justify lies told by trusted sources. This problem is made even worse by social recommendation algorithms which rect information, which can cause uncertainty and make dis- help amplify conspiracy theories and information confirming crediting information more difficult. The majority of technical one’s own biases due to companies’ efforts to optimize for clicks solutions have focused on classifying these extremes (fake and and watch time over individuals’ own values and public good. real), which leaves automatic assessment of uncertain, mixed As a result, more nuanced voices and facts are drowned out by veracity, and deeply contextual information difficult to assess. a continuous erosion of trust in better information sources. Most misinformation mitigation techniques assume that discrediting, A deeper problem that is left unaddressed in the technical filtering, or demoting low veracity information will help news research threads is what to do when information corrections, consumers make better information decisions. However, these whether done by an algorithm or journalist, do not work. negative results indicate that some news consumers, particularly Even if information is correctly discredited, consumers may extreme or conspiracy news consumers will not be helped. choose to ignore the correct information, due to distrust in the We argue that, given this background, technology solutions to combating misinformation should not simply seek facts or platform, algorithm, or organization providing the corrected discredit bad news sources, but instead use more subtle nudges information. This behavior is particularly prevalent among towards better information consumption. Repeated exposure to consumers with extreme or conspiratorial views [9]. If low such nudges can help promote trust in better information sources veracity information is filtered out or demoted, consumers may and also improve societal outcomes in the long run. In this article, become more extreme and distrust the contemporary media we will talk about technological solutions that can help us in developing such an approach, and introduce one such model platforms. The rise of alternative “free speech” platforms such called Trust Nudging. as Gab and Bitchute are examples of this [10]. Similarly, if Index Terms—misinformation, disinformation, decision sup- consumers perceive this filtering, demoting, or discrediting as port systems, information trust, nudge theory, recommendation partisan, distrust in information corrections can persist [11], [12], resulting in reduced trust for the news source, the platform/algorithm curating information and the fact-checking I.INTRODUCTION organization. Due to this distrust, solutions to correcting There are many useful and necessary paths to combat- misinformation can be ineffective for some consumers [9]. ing misinformation. These paths include technical methods In this paper, we begin to address this problem: How can to identify incorrect or misleading claims [1], [2], methods online media systems, if they are willing to, support trust arXiv:1911.05825v1 [cs.CY] 13 Nov 2019 to make correct information more easily available [3], and in higher quality sources? Specifically, we propose Trust methods to identify sources that disseminate incorrect infor- Nudging, a generic, trust-based recommendation model for mation [4], [5]. Some research pathways are non-technical, but improving the quality of news consumed. This proposed equally if not more important, as they address the underlying method is built on the concept of nudging, which provides issues and institutions that lead to the creation, dissemination, alternatives without forbidding any options or significantly and consumption of misinformation [6], [7]. There has also changing the economic incentives [13]. In essence, we would been significant growth in political fact-checking organiza- like to provide alternative sources of information to users at tions, including the fact-checking of news articles, social decision points without taking away any agency from them and media posts, and claims made by politicians [8]. without suppressing information. To do this, we provide subtle In all these paths, there are various underlying challenges. recommendations to readers in order to nudge them towards Overall, the misinformation problem is deeper than identifying news producers of objectively higher quality, but also have a what is “fake news.” While the dissemination of proven incor- chance of being trusted; thereby avoiding recommendations rect information is a real problem, confirming that specific that may not work or may break trust. We leverage news information is incorrect can be deeply contextual. This opens relationship graphs and the news already read by the consumer to approximate the trust of recommended news sources. Using Second, information trust plays a role in belief updating. a simulation, we show that this model can slowly increase the Information trust is a mix of an individual’s own judgment quality of news a consumer is reading, while not demanding of information and the trust in the source [25], [26]. When a substantial shift in trust or ideological beliefs. Furthermore, assessing trust, an individual may rely solely on their own we show that, as a side effect, this model lessens the partisan evaluation of the information, especially if the source is not extremity of the news being read. In addition to simulating trusted, in which case confirmation of the reader’s own beliefs this generic model, we outline different research threads that as well as other heuristics may play a large role [27]. For can help support this approach, as well as, different choice example, information that is compatible with a person’s current architectures that can support better information consumption. beliefs can be seen as more credible and stories that are Lastly, we discuss the benefits and potential drawbacks of this coherent may be easier to trust [27]. Many sources disseminat- type of recommendation method. ing misinformation have become quite good at targeting such heuristics [28]. Similarly, for trusted sources, the information II.RELATED WORK can be accepted as true without much critical evaluation. Trust A. Current Approaches to Misinformation Mitigation for sources is a complex concept as well, evaluated on multiple There have been many proposed technical solutions to axes, such as the alignment of the source’s values with the combating online misinformation. The vast majority of the reader or their perceived competence. technical solutions have been developed as detection systems, Over the past decade there has been an erosion of trust in which can filter out or discredit news that is of low veracity or media and political institutions [11], [12], which can material- provide fact-checking on the claims within the media. These ize as the polarization of trust in news outlets. If an algorithm solutions range widely in terms of technical methods used, recommends news from a high quality source that is initially including various types of machine learning models using the distrusted by the consumer, it is unlikely the consumer makes content in news articles and claims [1], [4], [14], deep neural a change. In the context of politics, a strongly partisan reader network models utilizing social features of shared news [15], may only trust sources closely aligned with their political source-level ranking models [3], and knowledge-graph models view. In this case, recommending an article from the opposite for fact-checking [2]. Many of these approaches have shown political camp is highly unlikely to work. Similarly, telling high accuracy in lab settings. Some approaches have also the reader of a conspiratorial news source to read an article shown robustness to concept drift and some adversarial at- from a neutral source is unlikely to yield any impact. As tacks [16]. both disinformation production and information trust become The assumption of detection-based misinformation solutions more politically polarized, methods that filter, block, demote, is that discrediting false or misleading information will help or discredit may be less effective, as they may be perceived consumers make fully informed decisions. However, there is as partisan paternalism. reason to believe that discrediting or filtering out bad infor- The decline of trust in long-standing news outlets is matched mation will not help some news consumers. First, discrediting with an increase in trust of information recommended on social information may be ignored by consumers with extreme or media [11], although as social media platforms become more conspiratorial views. As described in [9]: “A distinctive feature contemporary, this trust has also wavered. Partisan-based trust of conspiracy theories is their self-sealing quality. Conspiracy in information from social media is concerning as disinforma- theorists are not likely to be persuaded by an attempt to tion is often partisan and more prevalent on social media [29]. dispel their theories; they may even characterize that very Furthermore, a great deal of research and discussion shows that attempt as further proof of the conspiracy.” This negative social media recommendation systems further complicate this reaction is related to the “back-fire effect,” where the con- problem [30]. For example, Facebook’s news feed algorithm sumer’s beliefs become even stronger when information about has been said to “rewarded publishers for sensationalism, strongly-held beliefs is discredited [17], although this effect not accuracy or depth1.” As a result, news sources focused does not consistently occur [18]. Additionally, it has been on providing complete, neutral, and nuanced commentary on shown that discrediting misinformation does not always help factual events may end up being demoted in news feeds, reasoning, making the usefulness of discrediting information providing passive consumers with little opportunity to develop even more nuanced. Specifically, future reasoning can be trust for these high quality sources. influenced by misinformation even if that information has been Third, filtering out, blocking, or demoting bad information corrected or debunked (often called the continued-influence can be perceived as loss of agency or suppression of free effect) [19]. This lingering effect can appear as recalling the speech, which may increase polarization, particularly for those incorrect information from memory, even if it was corrected consumers with conspiratorial views. One very prominent ex- immediately [20], or maintaining strongly-held attitudes about ample of this is the rise of alternative media platforms such as the topic, even if beliefs are correctly updated [21], [22]. These Gab, Bitchute, and Voat [10], which harbor conspiracy theorist effects not only exist for explicit misinformation, but also and hyper-partisan information producers. These platforms more subtle, implied misinformation in news articles [23]. The self-proclaim that they have been created to promote free corrective effects of fact-checking can also be limited if they are given with unrelated contextual information [24]. 1www.wired.com/story/inside-facebook-mark-zuckerberg-2-years-of-hell/ speech rights that have been taken away from them through environmental nudges are unavoidable, Thaler and Sunstein the demonetization and removal from contemporary platforms argue that these environments should be purposely designed like Twitter and YouTube. While the movement of partisan and to make nudges beneficial [13]. conspiracy media from mainstream platforms to alternative There are situations where well-designed nudges may be platforms may stagnate misinformation flow to the wider- particularly helpful. People tend to make bad choices when public (which is still up for debate), it also creates even more a situation’s cost are not realized until later, the decision is extreme echo chambers, which can lead to radicalization [30]. difficult to assess, the choice being made is infrequent, or there These negative results do not necessarily invalidate the is no immediate feedback [13]. In these cases, a well-designed applicability of current approaches. The active engagement nudge can benefit the decision maker. of social scientists are needed to understand how and when Consuming online news cuts across many of these situ- employing such methods are necessary to reduce potential ations. Often consuming information has no immediate im- harm to individuals especially on social media. There is pact on the reader or sometimes the direct impact on the evidence that fact-checking efforts discourage politicians from individual consumer is never realized, as its impacts may be lying and can, under the right conditions, change consumer on a societal level. In some cases, there is a direct impact beliefs [8]. There is also evidence that automated methods help on the individual consumer, but it is delayed. For example, consumers determine if a news article is unreliable or biased, a news consumer may choose not to vaccinate their child particularly with feature-based explanations [31]. However, due to false information on vaccination safety. After some the benefits are not uniform: users who read or share news time, their child may catch a preventable disease due to this on social media benefit less from these type of explanations. misinformed decision. Because of this delayed and sometimes Lastly, there is evidence that flagging false news can reduce indirect feedback from information consumption, it may be false news sharing [32], which suggests that filtering out or difficult to translate the costs and benefits of news reading de- demoting maliciously false content can prevent its spread to cisions to the consumer. Additionally, as previously discussed the wider-public. Despite these positive results, there are also in Section II-A, humans use many mental shortcuts when ample negative results which should be addressed. This paper assessing information. The choice architecture of online news proposes one idea to begin addressing them. consumption often exacerbates the use of these shortcuts. One such example is the passive nature of scrolling through a social B. Nudges and Choice Architectures media feed. Rather than actively seeking information about Highly related to the method proposed in this paper is current events, users are passively consuming this information, the concept of nudging. The concept of nudging has been sometimes piecemeal, creating potentially incorrect factoids well studied in the field of Behavioral Economics. The idea when recalling reported events. This infinite-scrolling design was particularly popularized by Richard Thaler and Cass choice nudges users to consume passively. Sunstein’s 2008 book entitled Nudge: Improving Decisions There are many ways current news consumption systems, About Health, Wealth, and Happiness [13]. According to like social media, could be improved through better choice en- Thaler and Sunstein, a nudge is: vironments and nudges. In fact, many of these design changes are quite simple, but may oppose the current profit-driven “any aspect of the choice architecture that alters architectures of social media platforms. Examples include: people’s behavior in a predictable way without • Limiting infinite scrolling or auto-play features, to nudge forbidding any options or significantly changing users to more actively consume information, their economic incentives. To count as a mere • Setting the default news feed for new user accounts to nudge, the intervention must be easy and cheap to contain reliable and diverse information sources, utilizing avoid. Nudges are not mandates. Putting fruit at eye status quo bias [13], level counts as a nudge. Banning junk food does • Show diverse alternatives next to a given news article not.” in a news feed, providing an opportunity for the user to choose higher quality or diverse information, In other words, nudges are small changes in the choice ar- • Provide a portion of the body text with the title of an arti- chitecture that work with cognitive biases in decision making. cle in news feed, avoiding “implied misinformation. [23]’. The choice architecture is simply the environment in which In many of these systems, there is an additional layer a decision is being made, and that decision is nudged by in the form of recommendation or sorting, which explicitly this environment, even if it is not purposely designed to do influences what news is consumed (often much more explicitly so. Being nudged by a choice architecture is unavoidable. than one would consider a “nudge” to be). In a sense, these A classic example is the ordering of items in a list, where recommendation systems are filtering down the choices of the items atop the list are often selected more than any information to consume, in the same way Netflix filters down other item in the list, even if the ordering is arbitrary [33]. our movie choices based on what we previously watched. Another example is keeping the status quo. If something is However, while recommendation about what movie to watch set by default (i.e. the ringtone on a phone) it is unlikely is rather benign, recommending what information to consume to be changed [13]. Because these subtle, but influential, may not be. Information that is highly engaged with is not necessarily information of high quality [34]. Hence, removing news ecosystem. Media sources that hold the same ideological the engagement-based recommendation systems for news con- values, report on similar events, and write with similar styles sumption alone may be enough to nudge better consumption. are captured in these communities [35]. Hence, we expect sources which copy from each other at a high rate to be similar III.USING TRUSTFUL NUDGESIN NEWS to each other. Extending from this, we expect that sources RECOMMENDATIONS which are not necessarily directly connected, but near each Given the potential downfalls of discrediting partisan in- other in the network space (i.e. same community, peripheral formation and the complexity of information trust, we propose of the same community, etc.) should be alike in some regard. using “nudges” to help consumers make better consump- We assume that this similarity leads to a higher probability tion and sharing decisions. Specifically, we propose a trust- of trust for these sources than a randomly chosen source of based recommendation model for news quality called Trust higher quality. See Figure 1 for an example network. Nudging. The model’s goal is to provide subtle recommendations to readers in order to nudge them towards news producers Using this network we can model a consumer’s trust for a of objectively higher quality, without demanding a substantial source by placing that consumer’s reading profile as a node challenge to one’s beliefs. At a high level, suppose we are in the network. For example, given a consumer is subscribed given two partial orderings of information sources, one along to 5 sources on Facebook, we can use these 5 sources to a trust/belief axis and one along a nonpartisan quality axis. The approximate that consumers location in the network space (e.g. objective is given an article A, find an alternative article B that by averaging the vector representation of each of the 5 sources is higher in quality than A and has some proximity to A along in the network using an embedding method) and model trust as the trust dimension. The trust nudge is the act of providing the distance of consumer to any source in the network. If we to the user article A and B together in any point where they have a measure of quality for each news source (e.g. how many are making a decision, such as to share or to read. Users have times the source has published false news, etc.), we can ensure the choice to read or share one, both or none of the articles. our recommendation is of higher quality than the news the Both the quality dimension and the trust dimension can be consumer currently reads, but also ensure that the consumer approximated in many ways and be computed at different has a high chance of trusting that recommendation. Note, granularity. We present one approach in this paper. depending on the method of approximating the likelihood of To implement the Trust Nudging model, we need several trusting a recommendation, the recommended alternatives may pieces of information: not necessarily present very different or diverse points of view, 1) A relationship graph between news sources, which is especially if the initial article is on an extreme spectrum of used to approximate the likelihood that a user will trust unreliability. However, the main objective is to nudge the user a given news source. towards higher quality, rather than immediately recommend a 2) An approximate ground truth of both quality and po- gold standard article or source, which may be difficult for the litical leaning for each news source in the relationship user to accept. We contrast this approach with misinformation graph. This is used in both in the trust calculation and to detection methods that seek to discredit, demote, or filter out ensure recommendations are of higher quality than the information of lower quality. users’ current consumption. 3) A set of user reading profiles, which simply state what In some sense, this approach can also be contrasted with news sources a user reads or trusts. In real life this could traditional social recommendation algorithms that aim to only be the news a user is exposed to in his social media feed. maximize engagement, rather than quality. However, the goals The key to implementing the model is the news source of each recommendation algorithm are fundamentally differ- relationship graph. While theoretically this relationship graph ent. More importantly, inline with the concept of nudging, can represent a number of different relationships, such as text our approach is providing a recommendation as an alternative, similarity, topic similarity, or political leaning, the graph must not as the next step. It may be possible to implement both form a structure that relays the likelihood of consumer trust engagement-based recommendations (e.g. Facebook’s news across the sources. feed algorithm or YouTube’s up next feature) with the Trust One example of this structure is a news producer Content Nudging model, but this is left for future work. Sharing Network (CSN) constructed from a representative sample of news and media sources and articles published The quality ground truth required by the Trust Nudging them in a similar time frame. Several recent studies build on model can be built from various criteria, such as whether such datasets have shown that news sources often share (or they have published outright incorrect information in the past, copy) articles from each other either verbatim or in part [6], posted corrections to errors, provide transparency of financial [35], [36]. When these verbatim copies are formulated as interests, or distinguish between news and opinion. There are a network, they form meaningful communities of differing already organizations that provide this level of analysis, which sources in the news ecosystem [35], including communities of we will use in our proof-of-concept example [37]. While these conspiracy news, partisan news, and mainstream news. Each of source-level quality labels are a great start, such measures can these communities represent fairly homogeneous parts of the be further improved and made more granular. Fig. 1: News producer Content Sharing Network (CSN) used as the relationship graph in the Trust Nudging simulation based on data described in Section IV. Each node represents a news producer, each directed edge represents verbatim copy relationship (where A → B means B copies an article from A), the size of the node represents the nodes outdegree, and colors represent communities of sources computed using directed modularity.

IV. THE TRUST NUDGING MODELIN SIMULATION BuzzFeed). These labels include both measures of quality and To illustrate the Trust Nudging method, we construct a proof measures of political leaning. We extract and extend these of concept recommendation system based on real data. We labels for our simulation, discussed below. build each of the three sets of information from this data b) Relationship graph: While the Trust Nudging model set: the news relationship graph, the source entity, and the can be implemented with many different news relationship user entity. We then make recommendations based on the graphs, we use a content sharing network (CSN) in this trust nudge. We show how such a model may work based example (as discussed in Section III). Using our combined on simulations based on certain assumptions of user behavior dataset, we construct a CSN using the same method used and discuss the results. in [35]. Specifically, we compute a TF-IDF matrix of all a) Data: To ground our work on real data, we extract articles in the data set and compute the cosine similarity news article data from the NELA-GT-2018 dataset [37] and between each article vector pair (given that each article comes combine it with news article data we collected in the first from a different news source). For each pair of article vectors 9 months of 2019, giving us a total of 1,814,682 articles that have a cosine similarity greater or equal to 0.85, we extract from 271 sources over 19 months. The NELA-GT-2018 dataset them and order them by the timestamps provided with the is released with a set of source-level labels from multiple data set. This process creates a directed graph G = (V,E), organizations including independent and journalistic venues where V is a news source and E is a directed weighted (NewsGuard, Open Sources, Media Bias/Fact Check, Allsides, edge representing articles shared. Edges are directed towards Algorithm 1: Simulation algorithm s be a source in S, we define the quality score qs of s as

Data: u=User(Su) - the input user profile; a number in [0, 1] obtained via the aforementioned approach. S - the set of sources with quality score qs, leaning ls Similarly, the leaning ls of s is a number in [−1, 1]. v and embedded vectors vs; Additionally, we obtain a vector based representation s for T - the number of iterations T; each source s by embedding the s is a vector derived from L - user limited attention; the embedding CSN using node2vec [38] with 64-dimensional α - cost parameter; vectors. This representation captures the closeness, or similar- Result: Updated user profile u0 ity, of nodes in the CSN. initialization; d) User entity: Lastly, we need a set of users to nudge in t ← 0; the simulation. Let u user in the set of users U. User u has a set while t < T do Su of trusted sources, which has a maximum size specified by 0 a limited attention parameter L. We assume that a user has a s = argmins(t(s, u))|qs > qu); if s0 is not NULL then limited number of sources that it can attend to, or trust, at any moment. From set S we determine the quality score, leaning, if |Su| ≥ L then u 0 and CSN representation, q , l , v , respectively, associated drop source(u, {Su, s }); u u u end with u on their news consumption profile by averaging the else quality score, leaning, and CSN vectors of the sources in Su. accept source(u, s0); e) Trusted recommendation: Now that we have a news end relationship graph, a set of sources with ground truth, and update scores(u); a set of users with reading profiles, we can begin making end recommendations. A recommendation consists of suggesting 0 end a new source s to user u at a time t. The probability that 0 return u; user u will trust s depends on u’s profile. We denote it by 0 the conditional probability pu(s |Su). Our model considers both the leaning and the CSN representation of u and s0 to 0 compute pu(s |Su). More specifically, we compute the leaning publishers that copy articles (inferred by the timestamps). differential ∆lu,s0 as the normalized distance between lu and We normalize the weight of each edge in the network by ls0 , and we compute the source distance dus0 as the cosine the number of articles published in total by the source. The distance between vu and vs0 . We define the trust cost of end result is a near-verbatim content sharing network of recommending s0 to u as: news sources. This network can be found in Figure 1. After constructing the network (which naturally filters down both 0 0 t(s , u) = p (s |S ) = (1 − α)∆l + αd 0 (1) articles and sources), we have 102,879 pairs of articles and u u us 195 sources (which have copied or been copied from) over 19 Where α ∈ (0, 1) is a hyper-parameter to control the weight months. The constructed network can be found in Figure 1. of the source distance and the leaning in the function, ∆l = |lu−ls0 | c) Source entity: In order build our simulation, we must 2 , and dus0 = 1 − cos(vu, vs0 ). Concretely, we define know both the approximate quality of a news source and the the trust cost as a function of how dissimilar the new source political leaning. Both quality and political leaning can be is from the user’s profile with respect to their position in the extracted from the ground truth provided in the NELA-GT- CSN and their political leaning. 0 2018 dataset. Specifically, we determine the quality score of If the user accepts the recommendation, source s is added a source by normalizing the scores provided by NewsGuard. In to Su. If |Su| = L, the user drops one source at random chosen addition, we give a score of 0 to sources that have been flagged from the following distribution: with at least one of the following labels: by Open sources as fake, conspiracy, junksci, or unreliable or by Media Bias/Fact t(s, u) 0 pdrop(s) = P for si ∈ Su ∪ {s } Check as conspiracy, pseudoscience, or questionable source. t(si, u) The leaning score of a source is computed by averaging the Note that s0 may be dropped in this process, which means it scores from fields AllSides bias rating, Buzzfeed leaning, was rejected by the user. If s0 is not dropped, the user leaning Media Bias/Fact Check left bias or right bias, the resulting and quality score are updated. leaning score is normalized to the interval [−1, 1], where −1 f) Simulation of trusted recommendations over time: indicates a left-wing bias and 1 a right-wing bias. For the Given a user profile u, the simulation runs for T discrete few sources that do not have any labeled data in NELA-GT- time steps. At each time step t ≤ T , the model produces 2018, we estimate the quality score and leaning by averaging 0 a recommended source s meeting the criteria that qs0 > qu its neighbors quality score and leaning in the CSN. It can be requiring the least trust cost, thus: argued that this may be a noisy approximation of both quality and leaning, but for the purposes of our simulated example 0 these potentially noisy labels are fine. More specifically, for s = argmins(t(s, u)|qs > qu) 0 If the user’s set of sources |Su| ≥ L, one source in Su ∪ s change, allowing for higher quality (or more dissimilar to the is chosen and discarded (this can be thought of as the source original reading profile) sources to be recommended. It is clear being distrusted by the user or simply no longer being read that the jump from Freedom Bunker to Fox News is less costly by the user due to limited attention). Otherwise, s0 is accepted than the jump from Freedom Bunker to Politico. Similarly, the with probability 1 − t(s0, u). If s0 is accepted, u’s profile is jump from TheAntiMedia to RT is less costly than the jump updated by calculating the new means for qu, lu, and vu. This from TheAntiMedia to NPR. procedure is repeated for T iterations, when the updated profile Interestingly, as a side effect of increasing quality, the model of u is obtained. Notice that once qu = 1, no recommendation also lessens the partisan extremity of the news being read (with is made since the user has reached the maximum score possible the exception of User D, who already was not very extreme according to the model, we refer to the earliest t at which in political leaning, only quality). qu = 1 as the convergence point. Algorithm 1 outlines the b) Comparison to a trust-unconstrained model: The simulation procedure. Trust Nudging model is a trust-constrained model because the choice of recommendation is a function of the trust cost. A. Simulation Results A trust-unconstrained model would provide recommendations We ran the simulation on four synthetically developed user that are not necessarily functions of trust, or do not take profiles. The profiles were chosen to depict users with distinct trust directly into account when making decisions. A trust- characteristics with respect to political leaning and the types unconstrained model can be thought of as a recommendation of sources they trust (conspiracy, left-leaning, right-leaning, system that recommends gold standard news no matter the etc.). User profiles are shown in Table I. consumer. For comparison, we modified our model to become a) Pathways to high quality news: Figure 2 shows the trust-unconstrained by removing the trust function from the quality score (left side) and leaning (right side) trajectories source selection step and simply picking a source that incre- for each user. All users converge to a quality of 1, although at ments the overall quality score of the user. We then looked different times, the gray box indicates convergence to a quality at the progression of the trust cost for each iterations step of 1. Some of the sources along each pathway are emphasized, t comparing the constrained and unconstrained models. As their scores (quality and leaning) are annotated in parentheses seen in Figure 3, the trust cost for the unconstrained model in each figure. The ending profile and scores of users can be started at a much higher value than the constrained one, as time seen in Table II. Observing the simulation results, we are not passed, the trust cost started to reduce because the user had only able to see the final set of sources of users, but also the incorporated high quality source to their profile and became pathway taken by each of them in order to reach a high quality naturally closer to good sources in the CSN space. Even news consumption standard at each time step. though both models converge at approximately the same time, Importantly, the first recommendation made for each user is one may argue that it is unlikely for a user to accept drastic of higher quality than the user’s average reading profile, but is changes during the first recommendations since they are highly very similar to both their reading profile in the CSN space and unlikely to trust them. In the simulation model, the user is their reading profile in terms of the political leaning. While the repeatedly exposed to the recommendation, eventually leading quality is always higher than the user’s average reading quality, to acceptance. However, in real life, a user may begin to it may not be much higher depending on the extremity of lean- distrust the recommendation system or stop using the platform ing and connectedness of the reading profile in the CSN. For due to repeated and abrupt recommendations. example, User A is first recommended Freedom Bunker, which is a low quality (quality score of 0.3) right-wing (leaning score V. DISCUSSIONAND FUTURE WORK of 0.5) source, but this source is both higher quality (user’s In this paper, we proposed a trust-based news recommen- quality score of 0.075) and less extreme (user’s leaning score dation model which nudges consumers towards higher quality of 0.622) than the user’s reading profile. User A had a 96.7% without demanding a substantial challenge of one’s beliefs. chance of accepting that recommendation. Similarly, User D is The potential benefit of such a model is the ability to move recommended a low quality (quality score of 0.2) left-leaning extreme or conspiracy news consumers towards higher quality (leaning score of 0.1) source that is of higher quality than the information, a population which is hard to persuade. We user’s reading profile. User D had a 94.2% chance of accepting provide a proof of concept for the Trust Nudging model this recommendation. In both cases, the user’s overall quality through simulation. Using this simulation we show that the score improves after accepting the first recommendation. As model can slowly improve news consumption over time while expected, user’s with less extreme profiles have a higher recommending sources which the user is likely to trust. We chance of accepting the first recommendation and converge to also show that, as a side effect of improved quality, hyper- high quality quicker. User B had a 97.8% chance of accepting partisan news consumers eventually consume less partisan the first recommendation and User C had a 99.3% chance of news. Although conclusions that can be drawn from this accepting the first recommendation. simulation are limited, the purpose of this paper is to open up As time progresses, users’ likelihood of trusting a rec- a new technical research path for misinformation mitigation, ommended source will remain small. However, as the users namely using nudges in news algorithms rather than filtering, accept recommended sources, their reading profile will slowly demoting, and discrediting. Quality Score Political Leaning

Right bias Fox News (0.8) 1.00 Real Clear Politics (0.5) 1.0

0.75 Fortune (1.0) National Review (1.0) 0.8 Politico (1.0) 0.50 Real Clear Politics (1.0) Right leaning Freedom-Bunker (0.5) Politico (-0.1) 0.25 0.6 Fortune (0.5) User A 0.00 0.4 National Review (1.0) 0.25

Fox News (0.8) Left leaning Leaning of user 0.50 0.2 Quality Score of user 0.75

0.0 Freedom-Bunker (0.3) 1.00 Left bias 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 Time Time

Right bias 1.00 1.0

0.75 The Denver Post (1.0) 0.8 0.50 CBS News (1.0) Right leaning The Intercept (0.9) 0.25 0.6 User B 0.00 The Denver Post (-0.5) 0.4 CBS News (-0.5) Washington Monthly (1.0) 0.25

Left leaning Leaning of user 0.50 0.2 Quality Score of user 0.75

0.0 The Intercept (-0.8) 1.00 Washington Monthly (-0.8) Left bias 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 Time Time

Right bias 1.00 1.0 Real Clear Politics (0.5)

0.75 Fortune (0.5) Politico (1.0) 0.8 Fortune (1.0) 0.50 Daily Caller (1.0) Real Clear Politics (1.0) Right leaning Politico (-0.1) 0.25 0.6 User C 0.00 Daily Caller (0.7) 0.4 0.25

Left leaning Leaning of user 0.50 0.2 Quality Score of user 0.75

0.0 1.00 Left bias 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 Time Time

Right bias 1.00 1.0

0.75 BBC (1.0) NPR (1.0) 0.8 0.50 Right leaning RT (0.1) Politico (-0.1) 0.25 0.6 BBC (-0.2) 0.00 User D Politico (1.0) 0.4 0.25

TheAntiMedia (0.1) NPR (-0.2) Left leaning Leaning of user 0.50 0.2 RT (0.3) Quality Score of user 0.75

0.0 TheAntiMedia (0.2) 1.00 Left bias 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 Time Time Fig. 2: Recommendation pathways. On the left are the trajectories of the quality score of the user, some sources are emphasized along the pathway, enclosed in parentheses are the quality score of each source; the gray box indicates convergence. On the right are the leaning trajectories, enclosed in parentheses are the leaning score of each source. User Starting Quality Starting Leaning Starting Sources User A 0.075 0.622 Infowars, Newswars, Prison Planet, Veterans Today, Natural News User B 0.350 -0.854 Daily Kos, Shareblue, Bipartisan Report, Delaware Liberal, Addicting Info User C 0.524 1.0 Breitbart, Conservative TreeHouse, CNS News, The Epoch Times, Western Journal User D 0.098 -0.158 21st Century Wire, Mint Press News, Global Research, The Duran, Intellihub

TABLE I: Starting User Proﬁles used in simulation. Quality score is a number between 0 and 1, where 1 is the highest quality. For each user this is computed by taking the average quality score of the news sources they read. Leaning is a number between -1 and 1, where -1 is far left leaning and 1 is far right leaning. Again, this score is computed by averaging the political leaning of the users reading proﬁle.

User Ending Quality Ending Leaning Ending Sources Time User A 1.0 0.516 National Review, Real Clear Politics, The American Conservative, Politico, Fortune 54 User B 1.0 -0.500 CBS News, The New York Times, The Guardian, Washington Post, The Denver Post 14 User C 1.0 0.516 Real Clear Politics, National Review, Fortune, The American Conservative, Politico 20 User D 1.0 -0.250 FiveThirtyEight, Business Insider, NPR, The Hill, BBC 59

TABLE II: User Proﬁles After simulation. Quality score is a number between 0 and 1, where 1 is the highest quality. For each user this is computed by taking the average quality score of the news sources they read. Leaning is a number between -1 and 1, where -1 is far left leaning and 1 is far right leaning. Again, this score is computed by averaging the political leaning of the users reading proﬁle. Time is the number of iterations the user took to get to a quality of 1.0.

titles play a big role in disseminating misinformation. Many

0.6 readers will consume news passively and form opinions based constrained on the titles alone. In such cases, even the exposure to slightly 0.5 unconstrained different framing of information may create some amount of doubt in low quality information, especially since sources 0.4 are chosen based on potentially trusted sources. Moreover, 0.3 this type of model can create ﬁnancial incentive for sources to provide a better quality of information instead of models Trust cost 0.2 based on engagement which tend to favor more sensational

0.1 information and models based on gold standards that can be hard to define in some cases. In an effective implementation 0.0 of our method, many sources have a chance to get engagement 0 10 20 30 40 50 60 70 Time by providing better quality information. Fig. 3: Comparison of the trust cost for the constrained and Of course there are many unknowns about this type of nudg- the unconstrained models for User A. Notice the initially trust ing recommendation model that should be assessed. While in cost for the unconstrained model. This suggests that users are theory our simulated trust calculation makes sense, in actuality unlikely to accept the recommendations at the beginning. trust is likely much more complex. Other trust factors, such as user stances on specific topics, whether a friend had read or shared the news, or coherence of news story, could be There are many ways this model can be improved. For ex- accounted for, as long a quality measure is still held as the ample, rather than only defining trust as a function of similarity objective. It is likely these more granular trust factors are in the CSN, more granular relationship graphs can be used, consumer-dependent and can change across readers. Namely, a such as the similarity of writing style between sources. We can user’s trust may be modeled not only with respect to their news also model trust as a function of the difference in attitude, or consumption profile, but also in terms of how much influence stance, towards a topic, provided that certain sources are more they receive from their peers and their susceptibility to such, reliable or more similar when reporting on specific topics. thus adding new dimensions to the definition of trust. This Additionally, multiple news relationship graphs can be used increased trust complexity may produce different behavior at once through a multi-layer network. The value of these over time. For example, rather than eventually converging to various news graphs, in terms of representing consumer trust, the highest quality sources, there may be a point in which should be tested. consumers can no longer be nudged, hence plateauing the This model can also be improved through more complex quality we can achieve. User studies should be implemented objective functions. For example, rather than only operating on to assess interaction with the algorithm. quality, the model can also nudge towards view-point diversity. There are also some potentially negative consequences of Such a system may expose users to better or more diverse in- using this type of model, as news consumers will continue to formation even if it is not clicked. Previous research shows that be exposed to bad and incorrect information without seeing any warnings regarding its incorrectness. Hence, users will [16] B. D. Horne, J. Nrregaard, and S. Adalı, “Robust fake news detection continue to form incorrect beliefs that may be very hard to over time and attack,” ACM Transactions of Intelligent Systems Tech- nology, 2019. overcome. In addition, even if higher quality information is [17] B. Nyhan and J. Reifler, “When corrections fail: The persistence of available, users may never choose to read it as they do not political misperceptions,” Political Behavior, vol. 32, no. 2, pp. 303– find it engaging. The hope is that this exposure to low veracity 330, 2010. [18] T. Wood and E. Porter, “The elusive backfire effect: mass attitudes news will diminish as the consumers are nudged; however, it steadfast factual adherence,” Political Behavior, pp. 1–29, 2016. is unknown how long it will take consumers to gain trust for [19] H. M. Johnson and C. M. Seifert, “Sources of the continued influence ef- higher quality sources and when/if they will stop accepting fect: When misinformation in memory affects later inferences.” Journal of Experimental Psychology: Learning, Memory, and Cognition, vol. 20, nudges. Another potentially negative consequence is the newly no. 6, p. 1420, 1994. created incentive for malicious sources to game the system. It [20] U. K. Ecker, J. L. Hogan, and S. Lewandowsky, “Reminders and repeti- is possible that a smart malicious source can work to become a tion of misinformation: Helping or hindering its retraction?” Journal of Applied Research in Memory and Cognition, vol. 6, no. 2, pp. 185–192, ’middle quality’ source, which will be recommended to users 2017. who previously consumed extreme news. These malicious [21] B. Swire-Thompson, U. K. Ecker, S. Lewandowsky, and A. J. Berinsky, sources may even be able to control both the extreme news “They might be a liar but theyre my liar: Source evaluation and the prevalence of misinformation,” Political Psychology, 2019. being consumed and the middle quality news being recom- [22] E. Thorson, “Belief echoes: The persistent effects of corrected misinfor- mended. The robustness to adversarial manipulation should mation,” Political Communication, vol. 33, no. 3, pp. 460–480, 2016. be explicitly tested. Finally, questions of how this model can [23] P. R. Rich and M. S. Zaragoza, “The continued influence of implied and explicitly stated misinformation in news reports.” Journal of experimen- work in harmony with other policies and community standards tal psychology: learning, memory, and cognition, vol. 42, no. 1, p. 62, regarding unacceptable content requires additional research. 2016. [24] R. K. Garrett, E. C. Nisbet, and E. K. Lynch, “Undermining the REFERENCES corrective effects of media-based political fact checking? the role of [1] K. Popat, S. Mukherjee, J. Strotgen,¨ and G. Weikum, “Credibility contextual cues and na¨ıve theory,” Journal of Communication, vol. 63, assessment of textual claims on the web,” in Proceedings of ACM CIKM. no. 4, pp. 617–637, 2013. ACM, 2016, pp. 2173–2178. [25] S. Adalı, Modeling Trust Context in Networks. Springer Briefs, 2013. [2] N. Hassan, G. Zhang, F. Arslan, J. Caraballo, D. Jimenez, S. Gawsane, [26] B. Hilligoss and S. Y. Rieh, “Developing a unifying framework of S. Hasan, M. Joseph, A. Kulkarni, A. K. Nayak et al., “Claimbuster: the credibility assessment: Construct, heuristics and interaction in context,” first-ever end-to-end fact-checking system,” Proceedings of the VLDB Information Processing and Management, vol. 44, pp. 1467–1484, 2008. Endowment, vol. 10, no. 12, pp. 1945–1948, 2017. [27] S. Lewandowsky, U. K. Ecker, C. M. Seifert, N. Schwarz, and J. Cook, [3] A. Barron-Cedeno,´ I. Jaradat, G. Da San Martino, and P. Nakov, “Misinformation and its correction: Continued influence and successful “Proppy: Organizing the news based on their propagandistic content,” debiasing,” Psychological Science in the Public Interest, vol. 13, no. 3, Information Processing & Management, 2019. 2012. [4] B. D. Horne, W. Dron, S. Khedr, and S. Adalı, “Assessing the news [28] B. Horne and S. Adali, “This just in: Fake news packs a lot in title, uses landscape: A multi-module toolkit for evaluating the credibility of news,” similar, repetitive content in text body, more similar to satire than real in WWW Companion, 2018. news,” in NECO 2017: The Second International Workshop on News [5] R. Baly, G. Karadzhov, D. Alexandrov, J. Glass, and P. Nakov, “Pre- and Public Opinion at ICWSM, 2017. dicting factuality of reporting and bias of news media sources,” in [29] R. Faris, H. Roberts, B. Etling, N. Bourassa, E. Zuckerman, and Proceedings of 2018 EMNLP, 2018. Y. Benkler, “Partisanship, propaganda, and disinformation: Online media [6] K. Starbird, A. Arif, T. Wilson, K. Van Koevering, K. Yefimova, and and the 2016 us presidential election,” Berkman Klein Center Research D. Scarnecchia, “Ecosystem or echo-system? exploring content sharing Publication, vol. 6, 2017. across alternative media domains,” 2018. [30] M. H. Ribeiro, R. Ottoni, R. West, V. A. Almeida, and W. Meira, [7] K. STARBIRD, A. ARIF, and T. WILSON, “Disinformation as collab- “Auditing radicalization pathways on youtube,” arXiv preprint orative work: Surfacing the participatory nature of strategic information arXiv:1908.08313, 2019. operations,” 2019. [31] B. D. Horne, D. Nevo, J. O’Donovan, J.-H. Cho, and S. Adalı, “Rating [8] M. Amazeen, “Sometimes political fact-checking works. sometimes it reliability and bias in news articles: Does ai assistance help everyone?” doesnt. heres what can make the difference,” The Washington Post, in Proceedings of the International AAAI Conference on Web and Social vol. 3, 2015. Media, vol. 13, no. 01, 2019, pp. 247–256. [9] C. R. Sunstein and A. Vermeule, “Conspiracy theories: Causes and [32] P. Mena, “Cleaning up social media: The effect of warning labels on cures,” Journal of Political Philosophy, vol. 17, no. 2, pp. 202–227, likelihood of sharing false news on facebook,” Policy & Internet, 2019. 2009. [33] M. J. Salganik, P. S. Dodds, and D. J. Watts, “Experimental study of [10] S. Zannettou, B. Bradlyn, E. De Cristofaro, H. Kwak, M. Sirivianos, inequality and unpredictability in an artificial cultural market,” science, G. Stringini, and J. Blackburn, “What is gab: A bastion of free speech vol. 311, no. 5762, pp. 854–856, 2006. or an alt-right echo chamber,” in Companion Proceedings of the The [34] S. Vosoughi, D. Roy, and S. Aral, “The spread of true and false news Web Conference 2018, 2018, pp. 1007–1014. online,” Science, vol. 359, no. 6380, pp. 1146–1151, 2018. [11] J. Turcotte, C. York, J. Irving, R. M. Scholl, and R. J. Pingree, [35] B. D. Horne, J. Nørregaard, and S. Adalı, “Different spirals of sameness: “News recommendations from social media opinion leaders: Effects on A study of content sharing in mainstream and alternative media,” in media trust and information seeking,” Journal of Computer-Mediated Proceedings of the International AAAI Conference on Web and Social Communication, vol. 20, no. 5, pp. 520–535, 2015. Media, vol. 13, no. 01, 2019, pp. 257–266. [12] T. Hanitzsch, A. Van Dalen, and N. Steindl, “Caught in the nexus: A [36] Y. Benkler, R. Faris, and H. Roberts, Network propaganda: Manipula- comparative and longitudinal analysis of public trust in the press,” The tion, disinformation, and radicalization in American politics. Oxford International Journal of Press/Politics, vol. 23, no. 1, pp. 3–23, 2018. University Press, 2018. [13] R. H. Thaler and C. R. Sunstein, Nudge: Improving decisions about [37] J. Nørregaard, B. D. Horne, and S. Adalı, “Nela-gt-2018: A large multi- health, wealth, and happiness. Penguin, 2009. labelled news dataset for the study of misinformation in news articles,” [14] G. B. Guacho, S. Abdali, N. Shah, and E. E. Papalexakis, “Semi- in Proceedings of the International AAAI Conference on Web and Social supervised content-based detection of misinformation via tensor embed- Media, vol. 13, no. 01, 2019, pp. 630–638. dings,” in 2018 ASONAM. IEEE, 2018, pp. 322–325. [38] A. Grover and J. Leskovec, “node2vec: Scalable feature learning for [15] N. Ruchansky, S. Seo, and Y. Liu, “Csi: A hybrid deep model for fake networks,” in Proceedings of the 22nd KDD Conference. ACM, 2016, news detection,” in Proceedings of the 2017 ACM on Conference on pp. 855–864. Information and Knowledge Management. ACM, 2017, pp. 797–806.